GPU‐Accelerated Multiobjective PSO for Large‐Swarm Engineering Optimisation

ABSTRACT High‐quality multiobjective engineering optimisation requires large candidate populations to approximate dense Pareto fronts, yet CPU‐bound algorithms such as NSGA‐II become impractical for large populations. Although GPU hardware can relax this constraint, only algorithms whose internal structure permits data‐parallel execution can fully exploit it. We present a PyTorch‐native GPU MOPSO framework that extends the Coello Coello external‐archive design with linearly decaying inertia, boundary‐safe iterative archive pruning and periodic particle mutation, and we empirically characterise how swarm size, GPU throughput, Pareto‐front quality and run‐to‐run reproducibility interact across a range of engineering optimisation problems. The principal empirical finding is that GPU hardware makes it practical to run swarms of – particles and that these larger swarms produce substantially better engineering outcomes: GPU PSO‐5000 achieves 40% lower generational distance than CPU NSGA‐II‐100 on combustion engine calibration (, Wilcoxon signed‐rank test, 20 independent seeds), whereas GPU PSO‐2000 achieves the best hypervolume on a 30‐dimensional neural‐network surrogate in 2.5 s, which is 55% faster than CPU NSGA‐II and has lower run‐to‐run variance. A scalability study from to identifies a GPU throughput sweet spot at (2.5 μs per particle per iteration), where variance collapses relative to , enabling reproducible single‐run deployment without ensemble averaging. These gains arise from two compounding effects: GPU batch inference reduces per‐particle surrogate evaluation cost by over CPU, whereas PSO’s parallel update rule avoids the nondominated sorting bottleneck that prevents NSGA‐II from scaling beyond . A controlled GPU‐vs‐GPU comparison isolates this second effect, revealing a timing gap at attributable solely to algorithmic structure. Results generalise across a real‐world UCI Gas Turbine dataset, a 6D/4‐objective engine calibration problem and a fixed 60‐s wall‐clock budget protocol under which GPU PSO’s 5‐s convergence enables 12 independent seeds within a single NSGA‐II run budget.